Automatic grain segmentation in cross-polarized photomicrographs of sedimentary rocks using psychophysics inspired models

Abstract

The present work analyses the effect of three spatial filters, modelling the psychophysics data on brightness perception, when applied to a difficult computer vision problem. This pertains to the automatic grain-matrix segmentation of cross-polar images of sedimentary rock thin sections, gathered through petrographic microscopy. These three filters are assumed to be modelling three modes of vision, viz. vision-at-a-glance (VG, through Magnocellular or M channel in primate visual pathway), vision-in-strong-contrast mode (VSC, through Parvocellular or P channel of visual pathway), and vision-with-scrutiny (VS, through Magno-Parvochannels combined referred as the PM filter). The aforementioned biologically inspired filters related to brightness–contrast modelling are applied on the cross-polar microscopic images. Then binary segmentation is performed on the original, as well as the filtered, images using a k-means clustering algorithm which effectively segment the input images according to the homogeneity present. The binary segmented output when compared with our manually built grain-matrix segmented ground truth using suitable indices, forecasts noticeable difference in terms of the applied filter on them and the grain size distribution in that image, manifested sometimes by the resolution scale of capturing the image. Mostly for images with higher resolution of microscope lens while capturing, or for images displaying larger grain size, the PM filtered image has shown better grain-matrix segmentation, though the effect of filtering is not very pronounced for such images. On the other hand, for lower resolution of microscope lens for capturing the image or for smaller grain size featured images, the P filtered image has consistently shown more accurate grain-matrix aggregation result as also distinct improvement over those not filtered. The novelty of the approach lies in the application of a psychophysics model that attempts to mimic the manual segmentation performed by the trained eye of the expert in various visual modes. The proposed method achieves a reasonably higher accuracy in grain segmentation when compared with relevant existing methods especially while dealing with more difficult images containing many small-sized grains.